IBM MNLP IE at CASE 2021 Task 1: Multigranular and Multilingual Event Detection on Protest News

Page created by Donald Douglas
 
CONTINUE READING
IBM MNLP IE at CASE 2021 Task 1:
                   Multigranular and Multilingual Event Detection on Protest News

                                  Parul Awasthy, Jian Ni, Ken Barker, and Radu Florian
                                                    IBM Research AI
                                              Yorktown Heights, NY 10598
                                {awasthyp, nij, kjbarker, raduf}@us.ibm.com

                                    Abstract                                   • Subtask 2 - Sentence Classification: deter-
                                                                                 mine whether a sentence expresses informa-
                In this paper, we present the event detection                    tion about a past or ongoing event.
                models and systems we have developed for
                Multilingual Protest News Detection - Shared
                                                                               • Subtask 3 - Event Sentence Coreference Iden-
                Task 1 at CASE 2021. 1 The shared task
                has 4 subtasks which cover event detection                       tification: determine which event sentences
                at different granularity levels (from document                   refer to the same event.
                level to token level) and across multiple lan-
                guages (English, Hindi, Portuguese and Span-                   • Subtask 4 - Event Extraction: extract event
                ish). To handle data from multiple languages,                    triggers and the associated arguments from
                we use a multilingual transformer-based lan-                     event sentences.
                guage model (XLM-R) as the input text en-
                coder. We apply a variety of techniques and
                                                                               Event extraction on news has long been popu-
                build several transformer-based models that
                perform consistently well across all the sub-               lar, and benchmarks such as ACE (Walker et al.,
                tasks and languages. Our systems achieve an                 2006) and ERE (Song et al., 2015) annotate event
                average F1 score of 81.2. Out of thirteen                   triggers, arguments and coreference. Most pre-
                subtask-language tracks, our submissions rank               vious work has addressed these tasks separately.
                1st in nine and 2nd in four tracks.                         Hürriyetoğlu et al. (2020) also focused on detecting
                                                                            social-political events, but CASE 2021 has added
            1    Introduction                                               more subtasks and languages.
            Event detection aims to detect and extract useful                  CASE 2021 addresses event information ex-
            information about certain types of events from text.            traction at different granularity levels, from the
            It is an important information extraction task that             coarsest-grained document level to the finest-
            discovers and gathers knowledge about past and                  grained token level. The workshop enables par-
            ongoing events hidden in huge amounts of textual                ticipants to build models for these subtasks and
            data.                                                           compare similar methods across the subtasks.
               The CASE 2021 workshop (Hürriyetoğlu et al.,                  The task is multilingual, making it even more
            2021b) focuses on socio-political and crisis event              challenging. In a globally-connected era, infor-
            detection. The workshop defines 3 shared tasks.                 mation about events is available in many different
            In this paper we describe our models and systems                languages, so it is important to develop models
            developed for “Multilingual Protest News Detec-                 that can operate across the language barriers. The
            tion - Shared Task 1” (Hürriyetoğlu et al., 2021a).           common languages for all CASE Task 1 subtasks
            Shared task 1 in turn has 4 subtasks:                           are English, Spanish, and Portuguese. Hindi is an
                                                                            additional language for subtask 1. Some of these
                • Subtask 1 - Document Classification: deter-               languages are zero-shot (Hindi), or low resource
                  mine whether a news article (document) con-               (Portuguese and Spanish) for certain subtasks.
                  tains information about a past or ongoing                    In this paper, we describe our multilingual
                  event.                                                    transformer-based models and systems for each
               1
                 Distribution Statement “A” (Approved for Public Release,   of the subtasks. We describe the data for the sub-
            Distribution Unlimited)                                         tasks in section 2. We use XLM-R (Conneau et al.,

                                                                        138
Proceedings of the 4th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE) , pages 138–146
                                      August 5–6, 2021, ©2021 Association for Computational Linguistics
Task     Language        Train     Dev     Test       task 4, we use the entire Spanish and Portuguese
            English (en)      8392     932     2971       data as training for the multilingual model.
     1      Spanish (es)      800      200     250           For the final submissions, we use all the provided
           Portuguese (pt)    1190     297     372        data, and train various types of models (multilin-
             Hindi (hi)         -        -     268        gual, monolingual, weakly supervised, zero-shot)
            English (en)     20543     2282    1290       with details provided in the appropriate sections.
     2      Spanish (es)      2193     548     686
           Portuguese (pt)    946      236     1445         3   Multilingual Transformer-Based
            English (en)      476      120     100              Framework
     3      Spanish (es)        -       11      40        For all the subtasks we use transformer-based lan-
           Portuguese (pt)      -       21      40        guage models (Vaswani et al., 2017) as the in-
            English (en)      2565     681     311        put text encoder. Recent studies show that deep
     4      Spanish (es)      106        -     190        transformer-based language models, when pre-
           Portuguese (pt)     87        -     192        trained on a large text corpus, can achieve bet-
                                                          ter generalization performance and attain state-of-
Table 1: Number of examples in the train/dev/test sets.   the-art performance for many NLP tasks (Devlin
Subtasks 1 and 3 counts show number of documents,
                                                          et al., 2019; Liu et al., 2019; Conneau et al., 2020).
and subtasks 2 and 4 counts show number of sentences.
                                                          One key success of transformer-based models is
                                                          a multi-head self-attention mechanism that can
2020) as the input text encoder, described in sec-        model global dependencies between tokens in input
tion 3. For subtasks 1 (document classification)          and output sequences.
and 2 (sentence classification), we apply multilin-          Due to the multilingual nature of this shared task,
gual and monolingual text classifiers with different      we have applied several multilingual transformer-
window sizes (Sections 4 and 5). For subtask 3            based language models, including multilingual
(event sentence coreference identification), we use       BERT (mBERT) (Devlin et al., 2019), XLM-
a system with two modules: a classification module        RoBERTa (XLM-R) (Conneau et al., 2020), and
followed by a clustering module (section 6). For          multilingual BART (mBART) (Liu et al., 2020).
subtask 4 (event extraction), we apply a sequence         Our preliminary experiments showed that XLM-R
labeling approach and build both multilingual and         based models achieved better accuracy than other
monolingual models (section 7). We present the            models. Hence we decided to use XLM-R as the
final evaluation results in section 8. Our mod-           text encoder. We use HuggingFace’s pytorch im-
els have achieved consistently high performance           plementation of transformers (Wolf et al., 2019).
scores across all the subtasks and languages.                XLM-R was pre-trained with unlabeled
                                                          Wikipedia text and the CommonCrawl Corpus of
2     Data                                                100 languages. It uses the SentencePiece tokenizer
The data for this task has been created using the         (Kudo and Richardson, 2018) with a vocabulary
method described in Hürriyetoğlu et al. (2021). The     size of 250,000. Since XLM-R does not use
task is multilingual but the data distribution across     any cross-lingual resources, it belongs to the
languages is not the same. In all subtasks there is       unsupervised representation learning framework.
significantly more data for English than for Por-         For this work, we fine-tune the pre-trained XLM-R
tuguese and Spanish. There is no training data            model on a specific task by training all layers of
provided for Hindi.                                       the model.
   As there are no official train and development
                                                            4   Subtask 1: Document Classification
splits, we have created our own splits. The details
are summarized in Table 1. For most task-language         To detect protest events at the document level, the
pairs, we randomly select 80% or 90% of the pro-          problem can be formulated as a binary text clas-
vided data as the training data and keep the remain-      sification problem where a document is assigned
ing as the development data. Since there is much          label “1” if it contains one or more protest event(s)
less data for Spanish and Portuguese, for some sub-       and label “0” otherwise. Various models have been
tasks, such as subtask 3, we use the Spanish and          developed for text classification in general and also
Portuguese data for development only; and for sub-        for this particular task (Hürriyetoğlu et al., 2019).

                                                      139
Model              en-dev     es-dev   pt-dev                 guages, denoted by XLM-R (en), XLM-R (es),
   XLM-R (en)            91.7       72.1     82.3                  and XLM-R (pt).
   XLM-R (es)            85.4       71.9     83.9
   XLM-R (pt)            85.5       75.2     84.8             The results of various models on the develop-
 XLM-R (en+es+pt)        90.0       75.2     88.3            ment sets are shown in Table 2. We observe that:

                                                                • A monolingual XLM-R model trained with
Table 2: Macro F1 score on the development sets for
subtask 1 (document classification).                              one language can achieve good zero-shot per-
                                                                  formance on other languages. For example,
                                                                  XLM-R (en), trained with English data only,
In our approach we apply multilingual transformer-                achieves 72.1 and 82.3 F1 score on Spanish
based text classification models.                                 and Portuguese development sets. This is con-
                                                                  sistent with our observations for other infor-
4.1   XLM-R Based Text Classification Models                      mation extraction tasks such as relation extrac-
In our architecture, the input sequence (document)                tion (Ni et al., 2020).
is mapped to subword embeddings, and the embed-
dings are passed to multiple transformer layers. A              • Adding a small amount of training data from
special token is added to the beginning of the input              other languages, the multilingual model can
sequence. This BOS token is  for XLM-R.                        further improve the performance for those
The final hidden state of this token, hs , is used as             languages. For example, with ∼1k addi-
the summary representation of the whole sequence,                 tional training examples from Spanish and
which is passed to a softmax classification layer                 Portuguese, XLM-R (en+es+pt) improves the
that returns a probability distribution over the pos-             performance by 3.1 and 6.1 F1 points on
sible labels:                                                     the Spanish and Portuguese development sets,
                                                                  compared with XLM-R (en).
             p = softmax(Whs + b)                (1)
                                                             4.2   Final Submissions
   XLM-R has L = 24 transformer layers, with                 For English, Spanish and Portuguese, here are the
hidden state vector size H = 1024, number of                 three submissions we prepared for the evaluation:
attention heads A = 16, and 550M parameters. We
learn the model parameters using Adam (Kingma                S1: We trained five XLM-R based document clas-
and Ba, 2015), with a learning rate of 2e-5. We                  sification models initialized with different ran-
train the models for 5 epochs. Clock time was 90                 dom seeds using provided training data from
minutes to train a model with training data from all             all three languages (multilingual models). The
the languages on a single NVIDIA V100 GPU.                       final output for submission 1 is the majority
   The evaluation of subtask 1 is based on macro-                vote of the outputs of the five multilingual
F1 scores of the developed models on the test data               models.
in 4 languages: English, Spanish, Portuguese, and
Hindi. We are provided with training data in En-             S2: For this submission we also trained five
glish, Spanish and Portuguese, but not in Hindi.                 XLM-R based document classification mod-
   The sizes of the train/dev/test sets are shown in             els, but only using provided training data
Table 1. Note that English has much more training                from the target language (monolingual mod-
data (∼10k examples) than Spanish or Portuguese                  els). The final output is the majority vote of
(∼1k examples), while Hindi has no training data.                the outputs of the five monolingual models.
   We build two types of XLM-R based text classi-
fication models:                                             S3: The final output of this submission is the ma-
                                                                 jority vote of the outputs of the multilingual
   • multilingual model: a model is trained with                 models built in (1) and the monolingual mod-
     data from all three languages, denoted by                   els built in (2).
     XLM-R (en+es+pt);
                                                                For Hindi, there is no manually annotated train-
   • monolingual models: a separate model is                 ing data provided. We used training data from En-
     trained with data from each of the three lan-           glish, Spanish and Portuguese, and augmented the

                                                       140
Model           en-dev     es-dev   pt-dev            We prepared three submissions on the test data
      XLM-R (en)         89.2       78.0     82.2           for each language (English, Spanish, Portuguese),
      XLM-R (es)         84.4       86.4     80.1           similar to those described in section 4.2.
      XLM-R (pt)         83.2       82.2     85.1
    XLM-R (en+es+pt)     89.4       86.2     85.6           6   Subtask 3: Event Sentence Coreference
                                                                Identification
Table 3: Macro F1 score on the development sets for
subtask 2 (sentence classification).                    Typically, for the task of event coreference reso-
                                                        lution, events are defined by event triggers, and
                                                        are usually marked in a sentence. Two event trig-
data with machine translated training data from En-     gers are considered coreferent when they refer to
glish to Hindi (“weakly labeled” data). We trained      the same event. In this task, however, the gold
nine XLM-R based Hindi document classification          event triggers are not provided; the sentences are
models with the weakly labeled data, and the fi-        deemed coreferent, possibly, on the basis of any
nal outputs are the majority votes of these models      of the multiple triggers that occur in the sentences
(S1/S2/S3 is the majority vote of 5/7/9 of the mod-     being coreferent, or if the sentences are about the
els, respectively).                                     same general event that is occurring. Given a docu-
                                                        ment, this event coreference subtask aims to create
5     Subtask 2: Sentence Classification                clusters of coreferent sentences.
To detect protest events at the sentence level, one        There is good variety in the research for coref-
can also formulate the problem as a binary text         erence detection. Cattan et al. (2020) rely only on
classification problem where a sentence is assigned     raw text without access to triggers or entity men-
label “1” if it contains one or more protest event(s)   tions to build coreference systems. Barhom et al.
and label “0” otherwise. As for document clas-          (2019) do joint entity and event extraction using a
sification, we use XLM-R as the input text en-          feature-based approach. Yu et al. (2020) use trans-
coder. The difference is that for sentence classi-      formers to compute the event trigger and argument
fication, we set max seq length (a parameter of         representation for the task.
the model that specifies the maximum number of             Following the recent work on event coreference,
tokens in the input) to be 128; while for document      our system is comprised of two parts: the classi-
classification where the input text is longer, we set   fication module and the clustering module. The
max seq length to be 512 (for documents longer          classification module uses a binary classifier to
than 512 tokens, we truncate the documents and          make pair-wise binary decisions on whether two
only keep the first 512 tokens). We train the models    sentences are coreferent. Once all sentence pairs
for 10 epochs, taking 80 minutes to train a model       have been classified as coreferent or not, the clus-
with training data from all the languages on a single   tering module clusters the “closest” sentences with
NVIDIA V100 GPU.                                        each other with agglomerative clustering, using a
   For this subtask we are provided with training       certain threshold, a common approach for corefer-
data in English, Spanish and Portuguese, and eval-      ence detection (Yang et al. (2015); Choubey and
uation is on test data for all three languages. The     Huang (2017); Barhom et al. (2019)).
sizes of the train/development/test sets are shown         Agglomerative clustering is a popular technique
in Table 1.                                             for event or entity coreference resolution. At the be-
   As for document classification, we build two         ginning, all event mentions are assigned their own
types of XLM-R based sentence classification mod-       cluster. In each iteration, clusters are merged based
els: a multilingual model and monolingual models.       on the average inter-cluster link similarity scores
The results of these models on the development          over all mentions in each cluster. The merging pro-
sets are shown in Table 3. The observations are         cedure stops when the average link similarity falls
similar to the document classification task. The        below a threshold.
multilingual model trained with data from all three        Formally, given a document D with n sentences
languages achieves much better accuracy than a          {s1 , s2 , ..., sn }, our system follows the procedure
monolingual model on the development sets of            outlined in Algorithm 1 while training. The input
other languages that the monolingual model is not       to the algorithm is a document, and the output is a
trained on.                                             list of clusters of coreferent event sentences.

                                                      141
Algorithm 1: Event Coreference Training                             Model    en-dev    es-dev   pt-dev
   Input: D = {s1 , s2 , ..., sn }, threshold t                        S1       83.4      93.3     80.4
   Output: Clusters {c1 , c2 , ..., ck }                               S2       87.7      82.4     85.5
 1 Module Classify(D):
                                                                       S3       88.8      81.7     91.7
 2    for (si , sj ) ∈ D do
 3        Compute simi,j                                      Table 4: CoNLL F1 score on the development sets for
                                                              subtask 3: Event Coreference.
 4        SIM ← SIM ∪ simi,j
 5      return SIM
                                                          we have the score for all sentence pairs, we call
 6
                                                          the clustering module to create clusters using the
 7   Module Cluster(D, SIM ,t):
                                                          coreference scores as clustering similarity scores.
 8     for (si ) ∈ D do
                                                             We use XLM-R large pre-trained models. We
 9         Assign si to cluster ci
                                                          trained our system for 20 epochs with learning rate
10         Add ci → C
                                                          of 1e-5. We experimented with various thresholds
11      moreClusters = True                               and chose 0.65 as that gave the best performance
12      while moreClusters do                             on development set. It takes about 1 hour for the
13         moreClusters = False                           model to train on a single V100 GPU.
14         for (ci , cj ) ∈ C do
15             score=0                                        6.2    Final Submissions
16             for (sk ) ∈ ci do                          For the final submission to the shared task we ex-
17                   for (sl ) ∈ cj do                    plore variations of the approach outlined in 6.1.
18                        score+=simk,l                   They are:
19             if score > t then
                                                              S1: This is the multilingual model. To train this
20                  Merge ci and cj
                                                                  we translate the English training data to Span-
21                  Update C
                                                                  ish and Portuguese and train a model with
22                  moreClusters = True
                                                                  original English, translated Spanish and trans-
                                                                  lated Portuguese data. The original Spanish
23      return C
                                                                  and Portuguese data is used as the develop-
24                                                                ment set for model selection.

                                                              S2: This is the English-only model, trained on En-
6.1    Experiments                                                glish data. Spanish and Portuguese are zero-
The evaluation of the event coreference task is                   shot.
based on the CoNLL coref score (Pradhan et al.,
                                                              S3: This is an English-only coreference model
2014), which is the unweighted average of the F-
                                                                  where the event triggers and place and time
scores produced by the link-based MUC (Vilain
                                                                  arguments have been extracted using our sub-
et al., 1995), the mention-based B 3 (Bagga and
                                                                  task 4 models (section 7). These extracted to-
Baldwin, 1998), and the entity-based CEAFe (Luo,
                                                                  kens are then surrounded by markers of their
2005) metrics. As there is little Spanish and Por-
                                                                  type, such as , , etc. in
tuguese data, we use it as a held out development
                                                                  the sentence. The binary classifier is fed the
set.
                                                                  sentence representation.
   Our system uses XLM-R large pretrained model
to obtain token and sentence representations. Pairs             The performance of these techniques on the de-
of sentences are concatenated to each other along             velopment set is shown in table 4.
with the special begin-of-sentence token and sepa-
rator token as follows:                                       7     Subtask 4: Event Extraction
              BOS < si > SEP < sj >                       The event extraction subtask aims to extract
  We feed the BOS token representation to the             event trigger words that pertain to demonstrations,
binary classification layer to obtain a probabilistic     protests, political rallies, group clashes or armed
score of the two sentences being coreferent. Once         militancy, along with the participating arguments

                                                        142
Model     en-dev     es-dev     pt-dev
                                                                                    n
         S1        80.57        -          -                               Lt = −
                                                                                    X
                                                                                          log(P (lwi ))     (2)
         S2        80.25     64.09      69.67                                       i=1
         S3        80.87        -          -
                                                              7.1    Experiments
Table 5: CoNLL F1 score on the development sets for        The evaluation of the event extraction task is the
subtask 4: Event Extraction.                               CoNLL macro-F1 score. Since there is little Span-
                                                           ish and Portuguese data, we use it either as train
                                                           in our multilingual model or as a held out develop-
in such events. The arguments are to be extracted          ment set for our English-only model.
and classified as one of the following types: time,           For contextualized word embeddings, we use the
facility, organizer, participant, place or target of the   XLM-R large pretrained model. The dense layer
event.                                                     output size is same as its input size. We use the
                                                           out-of-the-box pre-trained transformer models, and
  Formally the Event Extraction task can be
                                                           fine-tune them with the event data, updating all
summarized as follows: given a sentence
                                                           layers with the standard XLM-R hyperparameters.
s = {w1 , w2 , .., wn } and an event label set
                                                           We ran 20 epochs with 5 seeds each, learning rate
T = {t1 , t2 ..., tj }, identify contiguous phrases
                                                           of 3 · 10−5 or 5 · 10−5 , and training batch sizes
(ws , ..., we ) such that l(ws , .., we ) ∈ T .
                                                           of 20. We choose the best model based on the
   Most previous work (Chen et al. (2015); Nguyen          performance on the development set. The system
et al. (2016); Nguyen and Grishman (2018)) for             took 30 minutes to train on a V100 GPU.
event extraction has treated event and argument
extraction as separate tasks. But some systems                7.2    Final Submission
(Li et al., 2013) treat the problem as structured             For the final submission to the shared task we ex-
prediction and train joint models for event triggers          plore the following variations:
and arguments. Lin et al. (2020) built a joint system
for many information extraction tasks including               S1: This is the multilingual model trained with
event trigger and arguments.                                      all of the English, Spanish and Portuguese
                                                                  training data. The development set is English
   Following the work of M’hamdi et al. (2019);                   only.
Awasthy et al. (2020), we treat event extraction as
a sequence labeling task. Our models are based                S2: This is the English-only model, trained on En-
on the stdBERT baseline in Awasthy et al. (2020),                 glish data. Spanish and Portuguese are zero-
though we extract triggers and arguments at the                   shot.
same time. We use the IOB2 encoding (Sang and
Veenstra, 1999) to represent the triggers and the             S3: This is an ensemble system that votes among
argument labels, where each token is labeled with                 the outputs of 5 different systems. The vot-
its label and an indicator of whether it starts or                ing criterion is the most frequent class. For
continues a label, or is outside the label boundary               example, if three of the five systems agree on
by using B-label, I-label and O respesctively.                    a label then that label is chosen as the final
                                                                  label.
   The sentence tokens are converted to token-level
contextualized embeddings {h1 , h2 , .., hn }. We          The results on development data are shown in table
pass these through a classification block that is com-     5. There is no score for S1 and S3 for es and pt as
prised of a dense linear hidden layer followed by a        all provided data was used to train the S1 model.
dropout layer, followed by a linear layer mapped
to the task label space that produces labels for each         8     Final Results and Discussion
token {l1 , l2 , .., ln }.                                 The final results of our submissions and rankings
   The parameters of the model are trained via cross       are shown in Table 6. Our systems achieved con-
entropy loss, a standard approach for transformer-         sistently high scores across all subtasks and lan-
based sequence labeling models (Devlin et al.,             guages.
2019). This is equivalent to minimizing the nega-             To recap, our S1 systems are multilingual models
tive log-likelihood of the true labels,                    trained on all three languages. S2 are monolingual

                                                        143
Our Scores            Best Competitor       Our
    Task - Language                                  S1     S2      S3              Score            Rank
    1 (Document Classification) - English           83.60 83.87 83.93               84.55             2
    1 (Document Classification) - Portuguese        82.77 84.00 83.88               82.43             1
    1 (Document Classification) - Spanish           73.86 77.27 74.46               73.01             1
    1 (Document Classification) - Hindi             78.17 77.76 78.53               78.77             2
    2 (Sentence Classification) - English           84.17 84.56 83.22               85.32             2
    2 (Sentence Classification) - Portuguese        88.08 84.87 88.47               87.00             1
    2 (Sentence Classification) - Spanish           88.61 87.59 88.37               85.17             1
    3 (Event Coreference) - English                 79.17 84.44 77.63               81.20             1
    3 (Event Coreference) - Portuguese              89.77 92.84 90.33               93.03             2
    3 (Event Coreference) - Spanish                 82.81 84.23 81.89               83.15             1
    4 (Event Extraction) - English                  75.95 77.27 78.11               73.53             1
    4 (Event Extraction) - Portuguese               73.24 69.21 71.5                68.14             1
    4 (Event Extraction) - Spanish                  66.20 62.02 66.05               62.21             1

Table 6: Final evaluation results and rankings across the subtasks and languages. Scores for subtasks 1 and 2 are
macro-average F1 ; subtask 3 are CoNLL average F1 ; subtask 4 are CoNLL macro-F1 . The ranks and best scores
are shared by the organizers. Bold score denotes the best score for the track.

models: for subtasks 1 and 2 they are language-             belling task, helps improve performance across
specific, but for subtasks 3 and 4 they are English-        languages. As there is much less training data
only. S3 is an ensemble system with voting for              for Spanish and Portuguese, pooling all languages
subtasks 1, 2 and 4, and an extra-feature system for        helps.
subtask 3. Among our three systems, the multilin-
gual models achieved the best scores in three tracks,       9   Conclusion
the monolingual models achieved the best scores in        In this paper, we presented the models and systems
six tracks, and the ensemble models achieved the          we developed for Multilingual Protest News Detec-
best scores in four tracks.                               tion - Shared Task 1 at CASE 2021. We explored
   For subtask 1 (document-level classification),         monolingual, multilingual, zero-shot and ensem-
the language-specific monolingual model (S2) per-         ble approaches and showed the results across the
forms better than the multilingual model (S1) for         subtasks and languages chosen for this shared task.
English, Portuguese and Spanish; while for subtask        Our systems achieved an average F1 score of 81.2,
2 (sentence-level classification), the multilingual       which is 2 F1 points higher than best score of other
model outperforms the language-specific mono-             participants on the shared task. Our submissions
lingual model for Portuguese and Spanish. This            ranked 1st in nine of the thirteen tracks, and ranked
shows that building multilingual models could be          2nd in the remaining four tracks.
better than building language-specific monolingual
models for finer-grained tasks.                             Acknowledgments and Disclaimer
   The monolingual English-only model (S2) per-           This research was developed with funding from
forms best on all three languages for subtask 3.          the Defense Advanced Research Projects Agency
This could be because the multilingual model (S1)         (DARPA) under Contract No. FA8750-19-C-0206.
here was trained with machine translated data.            The views, opinions and/or findings expressed are
Adding the trigger, time and place markers (S3) did       those of the author and should not be interpreted
not help, even when these features showed promise         as representing the official views or policies of the
on the development sets.                                  Department of Defense or the U.S. Government.
   The multilingual model (S1) does better for
Spanish and Portuguese on subtask 4. This is
consistent with our findings in Moon et al. (2019)          References
where training multilingual models for Named En-            Parul Awasthy, Tahira Naseem, Jian Ni, Taesun Moon,
tity Recognition, also a token-level sequence la-             and Radu Florian. 2020. Event presence predic-

                                                      144
tion helps trigger detection across languages. CoRR,       (CASE 2021), online. Association for Computational
  abs/2009.07188.                                            Linguistics (ACL).
Amit Bagga and Breck Baldwin. 1998. Algorithms for         Ali Hürriyetoğlu, Hristo Tanev, Vanni Zavarella, Jakub
 scoring coreference chains. In In The First Interna-        Piskorski, Reyyan Yeniterzi, and Erdem Yörük.
 tional Conference on Language Resources and Eval-           2021b. Challenges and applications of automated
 uation Workshop on Linguistics Coreference, pages           extraction of socio-political events from text (CASE
 563–566.                                                    2021): Workshop and shared task report.             In
                                                             Proceedings of the 4th Workshop on Challenges
Shany Barhom, Vered Shwartz, Alon Eirew, Michael             and Applications of Automated Extraction of Socio-
  Bugert, Nils Reimers, and Ido Dagan. 2019. Re-             political Events from Text (CASE 2021), online. As-
  visiting joint modeling of cross-document entity and       sociation for Computational Linguistics (ACL).
  event coreference resolution. In Proceedings of the
  57th Annual Meeting of the Association for Com-          Ali Hürriyetoğlu, Erdem Yörük, Deniz Yüret, Çağrı
  putational Linguistics, pages 4179–4189, Florence,         Yoltar, Burak Gürel, Fırat Duruşan, Osman Mutlu,
  Italy. Association for Computational Linguistics.          and Arda Akdemir. 2019. Overview of clef 2019
                                                             lab protestnews: Extracting protests from news
Arie Cattan, Alon Eirew, Gabriel Stanovsky, Mandar
                                                             in a cross-context setting. In Experimental IR
  Joshi, and Ido Dagan. 2020. Streamlining cross-
                                                             Meets Multilinguality, Multimodality, and Interac-
  document coreference resolution: Evaluation and
                                                             tion, pages 425–432, Cham. Springer International
  modeling.
                                                             Publishing.
Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and
  Jun Zhao. 2015. Event extraction via dynamic multi-      Ali Hürriyetoğlu, Vanni Zavarella, Hristo Tanev, Er-
  pooling convolutional neural networks. In Proceed-         dem Yörük, Ali Safaya, and Osman Mutlu. 2020.
  ings of the 53rd Annual Meeting of the Association         Automated extraction of socio-political events from
  for Computational Linguistics and the 7th Interna-         news (AESPEN): Workshop and shared task report.
  tional Joint Conference on Natural Language Pro-           In Proceedings of the Workshop on Automated Ex-
  cessing (Volume 1: Long Papers), pages 167–176,            traction of Socio-political Events from News 2020,
  Beijing, China. Association for Computational Lin-         pages 1–6, Marseille, France. European Language
  guistics.                                                  Resources Association (ELRA).

Prafulla Kumar Choubey and Ruihong Huang. 2017.            Ali Hürriyetoğlu, Erdem Yörük, Osman Mutlu, Fırat
  Event coreference resolution by iteratively unfold-        Duruşan, Çağrı Yoltar, Deniz Yüret, and Burak
  ing inter-dependencies among events. In Proceed-           Gürel. 2021.      Cross-Context News Corpus for
  ings of the 2017 Conference on Empirical Methods           Protest Event-Related Knowledge Base Construc-
  in Natural Language Processing, pages 2124–2133,           tion. Data Intelligence, pages 1–28.
  Copenhagen, Denmark. Association for Computa-
  tional Linguistics.                                      Diederik P. Kingma and Jimmy Ba. 2015. Adam: A
                                                             method for stochastic optimization. In Proceedings
Alexis Conneau, Kartikay Khandelwal, Naman Goyal,            of the 3rd International Conference on Learning
  Vishrav Chaudhary, Guillaume Wenzek, Francisco             Representations (ICLR), ICLR ’15.
  Guzmán, Edouard Grave, Myle Ott, Luke Zettle-
  moyer, and Veselin Stoyanov. 2020. Unsupervised          Taku Kudo and John Richardson. 2018. SentencePiece:
  cross-lingual representation learning at scale. In         A simple and language independent subword tok-
  Proceedings of the 58th Annual Meeting of the Asso-        enizer and detokenizer for neural text processing. In
  ciation for Computational Linguistics, pages 8440–         Proceedings of the 2018 Conference on Empirical
  8451, Online. Association for Computational Lin-           Methods in Natural Language Processing: System
  guistics.                                                  Demonstrations, pages 66–71, Brussels, Belgium.
                                                             Association for Computational Linguistics.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
   Kristina Toutanova. 2019. BERT: Pre-training of         Qi Li, Heng Ji, and Liang Huang. 2013. Joint event
   deep bidirectional transformers for language under-       extraction via structured prediction with global fea-
   standing. In Proceedings of the 2019 Conference           tures. In Proceedings of the 51st Annual Meeting of
   of the North American Chapter of the Association          the Association for Computational Linguistics (Vol-
   for Computational Linguistics: Human Language             ume 1: Long Papers), pages 73–82.
  Technologies, Volume 1 (Long and Short Papers),
   pages 4171–4186, Minneapolis, Minnesota. Associ-        Ying Lin, Heng Ji, F Huang, and L Wu. 2020. A
   ation for Computational Linguistics.                      joint neural model for information extraction with
                                                             global features. In Proc. The 58th Annual Meet-
Ali Hürriyetoğlu, Osman Mutlu, Farhana Ferdousi            ing of the Association for Computational Linguistics
  Liza, Erdem Yörük, Ritesh Kumar, and Shyam               (ACL2020).
  Ratan. 2021a. Multilingual protest news detection -
  shared task 1, CASE 2021. In Proceedings of the 4th      Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey
  Workshop on Challenges and Applications of Auto-           Edunov, Marjan Ghazvininejad, Mike Lewis, and
  mated Extraction of Socio-political Events from Text       Luke Zettlemoyer. 2020. Multilingual denoising

                                                     145
pre-training for neural machine translation. CoRR,          Representation, pages 89–98, Denver, Colorado. As-
  abs/2001.08210.                                             sociation for Computational Linguistics.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-         Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
  dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,               Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
  Luke Zettlemoyer, and Veselin Stoyanov. 2019.               Kaiser, and Illia Polosukhin. 2017. Attention is all
  Roberta: A robustly optimized BERT pretraining ap-          you need. In NIPS, pages 6000–6010.
  proach. CoRR, abs/1907.11692.
                                                            Marc Vilain, John Burger, John Aberdeen, Dennis Con-
Xiaoqiang Luo. 2005. On coreference resolution per-          nolly, and Lynette Hirschman. 1995. A model-
  formance metrics. In Proceedings of Human Lan-             theoretic coreference scoring scheme. In Proceed-
  guage Technology Conference and Conference on              ings of the 6th Conference on Message Understand-
  Empirical Methods in Natural Language Processing,          ing, MUC6 ’95, page 45–52, USA. Association for
  pages 25–32, Vancouver, British Columbia, Canada.          Computational Linguistics.
  Association for Computational Linguistics.
                                                            Christopher Walker, Stephanie Strassel, Julie Medero,
Meryem M’hamdi, Marjorie Freedman, and Jonathan               and Kazuaki Maeda. 2006. Ace 2005 multilin-
 May. 2019. Contextualized cross-lingual event trig-          gual training corpus. Linguistic Data Consortium,
 ger extraction with minimal resources. In Proceed-           Philadelphia, 57.
 ings of the 23rd Conference on Computational Nat-
 ural Language Learning (CoNLL), pages 656–665,             Thomas Wolf, Lysandre Debut, Victor Sanh, Julien
 Hong Kong, China. Association for Computational              Chaumond, Clement Delangue, Anthony Moi, Pier-
 Linguistics.                                                 ric Cistac, Tim Rault, Rémi Louf, Morgan Funtow-
                                                              icz, and Jamie Brew. 2019. Huggingface’s trans-
Taesun Moon, Parul Awasthy, Jian Ni, and Radu Flo-            formers: State-of-the-art natural language process-
  rian. 2019. Towards lingua franca named entity              ing. ArXiv, abs/1910.03771.
  recognition with BERT. CoRR, abs/1912.01389.
                                                            Bishan Yang, Claire Cardie, and Peter Frazier. 2015.
Thien Huu Nguyen, Kyunghyun Cho, and Ralph Gr-                A Hierarchical Distance-dependent Bayesian Model
  ishman. 2016. Joint event extraction via recurrent          for Event Coreference Resolution. Transactions
  neural networks. In Proceedings of the 2016 Con-            of the Association for Computational Linguistics,
  ference of the North American Chapter of the Asso-          3:517–528.
  ciation for Computational Linguistics: Human Lan-
  guage Technologies, pages 300–309.                        Xiaodong Yu, Wenpeng Yin, and Dan Roth. 2020.
                                                              Paired representation learning for event and entity
Thien Huu Nguyen and Ralph Grishman. 2018. Graph              coreference.
  convolutional networks with argument-aware pool-
  ing for event detection. In Thirty-second AAAI con-
  ference on artificial intelligence.
Jian Ni, Taesun Moon, Parul Awasthy, and Radu Flo-
   rian. 2020. Cross-lingual relation extraction with
   transformers. CoRR, abs/2010.08652.
Sameer Pradhan, Xiaoqiang Luo, Marta Recasens, Ed-
  uard Hovy, Vincent Ng, and Michael Strube. 2014.
  Scoring coreference partitions of predicted men-
  tions: A reference implementation. In Proceed-
  ings of the 52nd Annual Meeting of the Association
  for Computational Linguistics (Volume 2: Short Pa-
  pers), pages 30–35, Baltimore, Maryland. Associa-
  tion for Computational Linguistics.
Erik F. Tjong Kim Sang and Jorn Veenstra. 1999. Rep-
   resenting text chunks. In Proceedings of the Ninth
  Conference on European Chapter of the Associa-
   tion for Computational Linguistics, EACL ’99, page
  173–179, USA. Association for Computational Lin-
   guistics.
Zhiyi Song, Ann Bies, Stephanie Strassel, Tom Riese,
  Justin Mott, Joe Ellis, Jonathan Wright, Seth Kulick,
  Neville Ryant, and Xiaoyi Ma. 2015. From light
  to rich ERE: Annotation of entities, relations, and
  events. In Proceedings of the The 3rd Workshop on
  EVENTS: Definition, Detection, Coreference, and

                                                      146
You can also read